ABSTRACT As early detection and better treatment have increased cancer patient survival rates, the importance of protecting normal organs during radiation treatment is drawing more attention. Cancers in the thoracic region, which include lung, esophageal, thymus, mesothelioma and breast cancers, are among the most pervasive and deadly cancers. The protection of normal thoracic organs including lungs, heart, esophagus and spinal cord is critical in reducing long term toxicity in such cancers. To avoid excessively high radiation doses to such organs-at-risk (OARs), they are required to be correctly segmented from simulation computed tomography (CT) scans during radiation treatment planning to get an accurate dosage distribution. Despite tremendous effort into the development of semi- or fully-automatic segmentation solutions, current automated segmentation software, mostly using the atlas-based methods, has not yet reached the level of accuracy and robustness required for clinical usage. Therefore, in current practice, significant manual efforts are still required in the OAR segmentation process. Manual contouring suffers from inter- and intra-observer variability as well as institutional variability where different sites adopt distinct contouring atlases and labeling criteria and thus leads to inaccuracy and variability in OAR segmentation. When OARs are very close to the treatment target, segmentation errors as small as a few millimeters can have a statistically significant impact on dosimetry distribution and outcome. In addition, it is also costly and time consuming as it can take 1-2 hours of a clinicians? time to segment major thoracic organs due to the large number of axial slices required. The associated human efforts would significantly increase if adaptive radiation therapy (ART) is used as OARs from two or more simulation CT scans need to be segmented to adjust treatment plans. In recent years, the rapid development of deep learning methods has revolutionized many computer-vision areas and the adoption of deep learning in medical applications has shown great success. Based on a deep-learning-based algorithm we developed that achieved better-than-human performance and ranked 1st in 2017 American Association of Physicist in Medicine Thoracic Auto-segmentation Challenge, a thoracic OAR auto-segmentation product will be developed in this project with the two aims: 1) improve and validate the deep-learning-based automatic thoracic organ segmentation algorithm on a larger clinical data set, and 2) incorporate this algorithm into a preliminary product that fits into the clinical workflow. With this product, the segmentation accuracy can be improved, leading to more robust treatment plans in protecting normal organs and improved long term patient outcome. Furthermore, the time and cost of radiation treatment planning can be greatly reduced, contributing to a more affordable cancer treatment and reduced healthcare burden.